Please Wait...

Accuracy of BMAS Hippocampus Segmentation Using the Harmonized Hippocampal Protocol

F. Roche, J. Schaerer, S. Gouttard, A. Istace, B. Belaroussi, HJ. Yu, L. Bracoud, C. Pachai , C. DeCarli and the Alzheimer's Disease Neuroimaging Initiative

1 BioClinica, Lyon, France & Newtown, PA, USA - email: joel.schaerer@bioclinica.com  2 University of California-Davis, Alzheimer's Disease Center, Sacramento, CA, USA

INTRODUCTION

  • Hippocampal volume (HCV) measured with MRI has been widely used as a key biomarker for both improving subject selection and monitoring treatment efficacy in Alzheimer’s Disease (AD) studies. However various hippocampal protocols exist in the literature, each including a different set of subfields and sub-structures, potentially leading to confusion and additional complexity for direct comparison and consistency in reporting study results.
  • The main goal of the Hippocampal Protocol (www.hippocampal-protocol.net) was to harmonize the multiple existing hippocampal protocols in order to define a standard protocol. Once that protocol was defined, expert tracers manually outlined hippocampal regions on sample ADNI cases. 100 expert hippocampus segmentations were made available as of January 2014. 35 more cases are expected to be released shortly [1].
  • The purpose of this work was to evaluate the accuracy of the BioClinica Multi-Atlas Segmentation (BMAS) algorithm - which had been previously validated using independent and in-house ADNI-based atlases [2-4] - with respect to the newly released hippocampal structures as defined by the Harmonized Hippocampal Protocol.

METHODS

Population

  • The Harmonized Hippocampal Protocol was based on 100 subjects from the ADNI-1 database. This cohort included 37 AD, 34 Mild Cognitive Impairment (MCI) and 29 Normal Controls (NC) subjects. Demographic data are reported in Table 1.
  • All subjects underwent MRI examinations composed of a high resolution 3DT1 sequence in compliance with the ADNI-1 imaging protocol, using MP-RAGE (Siemens), 3D TFE (Philips) and 3D Fast SPGR (General Electric) pulse sequences.
  • Images were re-oriented along the AC-PC line for the purpose of the Harmonized Protocol.
  • Images were manually contoured by expert tracers according to the Harmonized Hippocampal protocol.
  • The whole set of subjects was used as atlases for the BioClinica Multi-Atlas Segmentation algorithm (BMAS) [4].
  • Hippocampal segmentations were automatically computed with the BMAS algorithm in a "leave-one out" fashion to evaluate the accuracy of the proposed segmentation method with respect to this new standard.

Image analysis

Pre-processing

  • All atlases were registered to a common atlas space using an affine registration to improve computation time.

Multi-atlas segmentation

  • For a given subject among the 100 available cases, the 3DT1 sequence was affinely registered to the common atlas space.
  • Rectangular ROIs were automatically extracted around each hippocampus. All of the subsequent image processing was performed over those ROIs for improved segmentation accuracy.
  • The other 99 atlases were registered to the left-out subject using a non-linear registration method based on diffeomorphic demons.
  • The resulting segmentation results were fused using the MALF method [5], which weights atlases according to their local similarity to the subject.

Refinement and pooling

  • The resulting segmentation was subsequently refined by combining tissue-based classification and local intensity constraints.
  • An overview of the segmentation process is provided in Figure 1.

Statistical analysis

  • Segmentation accuracy was assessed using the Dice similarity index, signed and unsigned relative volume error (RVE) with the provided hippocampal segmentation as a reference.
  • Pearson correlation was also calculated, between computed and reference HCV values.
  • Finally, sensitivity of the manual (reference) and automatic HCV quantification methods in distinguishing AD, MCI and NC groups was measured by assessing group separation with the area under the ROC curve (AUC).

RESULTS & CONCLUSIONS

  • As detailed in Table 2, accuracy of the proposed BMAS segmentation was Dice=86.6% ± 1.7 and the related errors corresponding to mean hippocampal volumes (right + left) were RVE=4.5% ± 3.3 and signed RVE=2.8% ± 4.8.
  • Harmonized BMAS hippocampus segmentation results were highly correlated to the newly introduced harmonized hippocampus segmentation (r=0.96 with p<0.001).
  • Similar ability for subject group separation was observed for the reference and Harmonized BMAS hippocampus segmentations (see Table 2).
  • This work demonstrated that BMAS could be successfully implemented using the Harmonized Protocol atlases and provided sufficient accuracy. It is noteworthy that manual and automated segmentations lead to similar group dissociation ability on this small sample of ADNI1 subjects.

REFERENCES

  1. Providing standardized labels of the EADC-ADNI Harmonized Hippocampal Protocol for Automated Algorithm Training, Boccardi et al., CTAD 2013
  2. Reproducibility of Intracranial and Hippocampal Volume Quantification at 1.5T and 3T MRI – Application to ADNI I, Roche et al., AAIC 2013
  3. Comparison of Manual and Automated Multi-Atlas Hippocampal Volume Measurements in the UC Davis Alzheimer’s Disease Center Patient Cohort, Schaerer et al., AAIC 2013
  4. Multi-Atlas Hippocampus Segmentation Refined with Intensity-based Tissue Classification, Belaroussi et al., AAIC 2012
  5. Multi-Atlas Segmentation with Joint Label Fusion, Wang et al., Pattern Analysis and Machine Intelligence, March 2013

     

LEARN MORE OR SPEAK WITH OUR EXPERTS

CONTACT US
Leader in Clinical Trial
Management Solutions

Successful clinical trials require the ability to see key details and uncover hidden insights. Bioclinica utilizes science and technology to bring clarity to clinical trials, helping companies to develop new life-improving therapies more efficiently and safely.

Dr Joby John chairing a panel discussion tomorrow 8am at #DIA2019 on ways to improve connectedness of clinical data… https://t.co/BPzi1geKsv
Bioclinica (14 hours ago)
Meet the Bioclinica team at DIA Booth 1831. #DIA2019 https://t.co/pvvdiqJz32 https://t.co/tx4PLDIrMm
Bioclinica (16 hours ago)
We’re at #DIA2019. Visit us at booth #1831 to learn more about our rapidly growing services and meet face to face… https://t.co/rK7V1Aclov
Bioclinica (19 hours ago)
Learn about When EDC is not enough: Automating multi-country data collection and complete workflows at #DIA2019! Ja… https://t.co/zJfytNpz82
Bioclinica (1 week ago)
In San Diego for #DIA2019? Visit us at booth #1831 and meet face to face with our team to learn more about our rapi… https://t.co/T4U4oNXlxo
Bioclinica (2 weeks ago)
We're celebrating International #NASH Day to help raise awareness about non-alcoholic steatohepatitis & fatty liver… https://t.co/16PfGbzeaS
Bioclinica (2 weeks ago)

Latest Blogs:

Medical Imaging of Non-Alcoholic Fatty Liver Diseases in Clinical Trials
EDC Provider Selection
Site Payments Solutions
Nonalcoholic steatohepatitis (NASH) development programs
Dr. Joby John
Transformational Trends in Investigator Site Payments
Clinical Supply Chain Success: Leveraging Technology
Blood pressure monitor
Prescription box